PickleRick’s point about unstructured data is the part I’d keep in the design. Start with fields you can identify reliably, then inspect the actual prompts after replacement. A clean looking dashboard does not tell you what reached the model. I’m building Sether, a free MIT TypeScript and Python library for local supported pattern replacement. It also offers stable aliases for registered originals and local restoration, which may help when the answer needs usable values back. Splunk auditing and application policy would still govern the wider flow. Do your outputs need restoration, or is permanent masking enough? Docs: https://sether.raevenmarket.com.ng/docs
... View more